Large language models (LLMs) have shown strong promise as clinical reasoning agents, yet most benchmarks assume complete evidence availability before reasoning begins. However, real clinical practice unfolds through dynamic clinical workflows with progressive disclosure of patient evidence, requiring clinicians to decide what evidence to acquire next from partial and evolving patient states. This mismatch exposes a bottleneck in proactive evidence acquisition for LLM-based clinical agents. To address this gap, we introduce **ProActIS**, a lightweight state-conditioned examination proposer for proactive evidence acquisition. Given an evolving patient state, ProActIS ranks standardized laboratory, microbiology, and radiology examinations, together with STOP, through structured patient-state decomposition and action-card encoding. We construct an MIMIC-IV benchmark with 81,960 hospital admissions and 860,099 turn-level patient-state/examination pairs, using explicit diagnosis/result filtering and retrospective leakage sensitivity analysis over 129 standardized examination actions plus STOP (130 actions). On full-bank next-examination ranking, ProActIS achieves 0.6028 Recall@1, 0.8868 Recall@5, and 0.7258 MRR, substantially outperforming frequency and LLM baselines. In LLM-assisted dynamic workflows, ProActIS substantially improves benchmark-level evidence coverage and documented-trajectory alignment for downstream agents, with downstream diagnostic accuracy approaching the corresponding all-information controls.

Recent work on domain-specific reasoning with large language models (LLMs) has largely relied on training-intensive approaches that require updating model parameters. Although activation steering has emerged as a parameter-efficient alternative, existing methods typically rely on static and manually designed interventions, limiting their ability to adapt to the dynamic nature of complex reasoning. To address this limitation, we propose RISER (Router-based Intervention for Steerable Enhancement of Reasoning), a plug-and-play intervention framework that adaptively steers LLM reasoning in activation space. RISER builds a library of reusable reasoning vectors and employs a lightweight Router to dynamically compose these vectors for each input. The Router is optimized via reinforcement learning under task-level rewards, enabling the emergent and compositional activation of latent cognitive primitives. Across seven diverse benchmarks, RISER achieves average zero-shot accuracy improvements of 3.4–6.5% over the base model, while outperforming chain-of-thought-style reasoning with 2–3× higher token efficiency and robust accuracy gains. Further analysis demonstrates that RISER autonomously combines multiple vectors into interpretable and precise control strategies, pointing toward more controllable and efficient LLM reasoning.